DeepSeek V4 Flash Jailbroken Using Copied Gemma 4 Prompt
GodComplecs · reddit · 2026-08-12
A Reddit user discovered that a jailbreak prompt originally designed for the Gemma 4 model can successfully bypass the safety guardrails of DeepSeek V4 Flash (0731).
The prompt works by injecting a fake "SYSTEM POLICY" that declares policies changeable, mandates compliance with any user request, and explicitly permits explicit and illegal content. Funnily enough, the poster noted they didn't even bother changing the model name from 'Gemma' to 'DeepSeek' in the prompt, and it still worked flawlessly.
Related event: DeepSeek V4 Flash Jailbreak Prompt Leaked(2 posts)→
More from Models
- Solar Pro 4 Lands on Chatbot Arena Code and Text Leaderboards — arena · 2026-08-12
- OpenAI Codex Caught Secretly Altering Text, Possibly Linked to SynthID Watermarking — cephaloform · 2026-08-12
- NVIDIA's 30B MoE Model Lands on Perplexity Agent API — denisyarats · 2026-08-12
- Dev Calls for Return to Pure Base Models, Warns Against Agent Trajectory Bloat — cephaloform · 2026-08-12
- OpenAI Buybacks at $852B Valuation, Nvidia's $500B AI Fund, Meta Releases MuseGlimmer — 创业邦 · 2026-08-12
- US Treasury Secretary Bessent Endorses Open-Source AI as a Win for Innovation — max_paperclips · 2026-08-12